Edit model card

InCaseLawBERT

Model and tokenizer files for the InCaseLawBERT model from the paper Pre-training Transformers on Indian Legal Text.

Training Data

For building the pre-training corpus of Indian legal text, we collected a large corpus of case documents from the Indian Supreme Court and many High Courts of India. The court cases in our dataset range from 1950 to 2019, and belong to all legal domains, such as Civil, Criminal, Constitutional, and so on. In total, our dataset contains around 5.4 million Indian legal documents (all in the English language). The raw text corpus size is around 27 GB.

Training Setup

This model is initialized with the Legal-BERT model from the paper When does pretraining help?: assessing self-supervised learning for law and the CaseHOLD dataset of 53,000+ legal holdings. In our work, we refer to this model as CaseLawBERT, and our re-trained model as InCaseLawBERT. We further train this model on our data for 300K steps on the Masked Language Modeling (MLM) and Next Sentence Prediction (NSP) tasks.

Model Overview

This model uses the same tokenizer as CaseLawBERT. This model has the same configuration as the bert-base-uncased model: 12 hidden layers, 768 hidden dimensionality, 12 attention heads, ~110M parameters.

Usage

Using the model to get embeddings/representations for a piece of text

from transformers import AutoTokenizer, AutoModel
tokenizer = AutoTokenizer.from_pretrained("law-ai/InCaseLawBERT")
text = "Replace this string with yours"
encoded_input = tokenizer(text, return_tensors="pt")
model = AutoModel.from_pretrained("law-ai/InCaseLawBERT")
output = model(**encoded_input)
last_hidden_state = output.last_hidden_state

Fine-tuning Results

We have fine-tuned all pre-trained models on 3 legal tasks with Indian datasets:

  • Legal Statute Identification (ILSI Dataset)[Multi-label Text Classification]: Identifying relevant statutes (law articles) based on the facts of a court case
  • Semantic Segmentation (ISS Dataset)[Sentence Tagging]: Segmenting the document into 7 functional parts (semantic segments) such as Facts, Arguments, etc.
  • Court Judgment Prediction (ILDC Dataset)[Binary Text Classification]: Predicting whether the claims/petitions of a court case will be accepted/rejected

InCaseLawBERT performs close to CaseLawBERT across the three tasks, but not as good as InLegalBERT. For details, see our paper.

Citation

@article{paul-2022-pretraining,
  doi = {10.48550/ARXIV.2209.06049},
  url = {https://arxiv.org/abs/2209.06049},
  author = {Paul, Shounak and Mandal, Arpan and Goyal, Pawan and Ghosh, Saptarshi},
  title = {Pre-training Transformers on Indian Legal Text},
  publisher = {arXiv}, 
  year = {2022},
  copyright = {Creative Commons Attribution 4.0 International}
}

About Us

We are a group of researchers from the Department of Computer Science and Technology, Indian Insitute of Technology, Kharagpur. Our research interests are primarily ML and NLP applications for the legal domain, with a special focus on the challenges and oppurtunites for the Indian legal scenario. We have, and are currently working on several legal tasks such as:

  • named entity recognition, summarization of legal documents
  • semantic segmentation of legal documents
  • legal statute identification from facts, court judgment prediction
  • legal document matching You can find our publicly available codes and datasets here.
Downloads last month
19
Hosted inference API
Fill-Mask
Examples
Examples
Mask token: [MASK]
This model can be loaded on the Inference API on-demand.